Efficient subband coding of speech with optimized uniform DFT filter banks

Abstract

Uniform DFT filter banks (FB) offer reduced implementation complexity of subband coders (SBC) as compared with QMF-based SBC. However, because known design techniques aim at minimizing the overall response error of the FB, using either deterministic [V. K. Jain and R. E. Crochiere, Proc. ICASSP 83, 228–231 (1983)] or statistical [A. Dembo and D. Malah, IEEE Trans. Acoust. Speech Signal Process. ASSP-36, 328–341 (1988)] error measures, the performance of DFT-based SBC was found to be subjectively inferior. A new approach is presented for DFT-FB design, which is based on a frequency-doman distortion function in which different weights are applied to the different error components (including interband aliasing and leakage of both signal and quantization noise) thus allowing an optimized design for SBC. The minimization of the proposed distortion function results in two sets of linear equations that are solved iteratively to obtain the optimal analysis and synthesis prototype filters. A 16-band DFT-based SBC operating at 16 kb/s with the designed filters was found in simulations to have subjective and objective performance similar to a QMF-based SBC at the same rate, but with less than half of the computations.